Automated Segmentation of In Situ X-ray Microtomography of Progressive Damage in Advanced Composites via Deep Learning

We present here the development and evaluation of a deep learning (artificial intelligence)-based computer vision machine to automate segmentation of multiclass progressive matrix damage across micro and mesoscales in aerospace-grade advanced composite laminates visualized in 4D via nondestructive in situ mechanical testing coupled with synchrotron radiation computed tomography (SRCT). Leveraging tens of thousands of manually-/human-annotated SRCT tomograms (i.e., 2D virtual cross-sectional slices) encompassing two different aerospace-grade advanced composite laminate systems (standard-thickness-ply and thin-ply) that were SRCT-scanned while under progressive tensile loading, we teach a fully convolutional neural network machine to segment complex polymer matrix damage mechanisms according to their host ply, replacing ~10 hours of trained human labor per scan segmentation (~2000 tomograms per scan) with negligible time to configure the trained machine data-processing pipeline. Evaluating qualitatively and quantitatively the segmented tomograms independently in 2D, as well as collectively in 3D scans, we demonstrate good agreement between the state-of-the-art human-based region growing (semi-manual) method and machine-based segmentation results, summarized by test set macro-averages of the following common classification/segmentation performance metrics: 79% for F1 score (harmonic mean of precision and recall) and 67% for intersection over union (IoU) score. Moreover, 2D inspection of segmented damage within tomograms reveals that F1 and IoU scores actually underrate machine performance due to a nontrivial degree of human (used as ground truth) segmentation error, as the machine is found to regularly exceed the human (resulting in F1 and IoU score penalties) by discovering new damage instances, augmenting existing diffuse segmentations, and extending segmentations to image artifact-prone specimen edges. Consequently, we discover that deep learning-based segmentation successfully and efficiently characterizes sparse (<<1% of scan volume), extremely complex 3D damage states within SRCT datasets, resolving an intractable computer vision challenge (as viewed through the lens of traditionally programmed automation) and establishing these high-throughput tools as promising candidates to accelerate understanding of basic structure-property relationships in traditional and next-generation advanced composite materials.